DeepSeek V4 Flash Vision Exp: Preview Release Facts
DeepSeek released the experimental DeepSeek-V4-Flash-Vision-Exp model on August 21, 2026.
Release Details and Access Method
The original announcement for this model occurred on August 21, 2026. On that date, DeepSeek stated that the new multimodal vision understanding model, DeepSeek-V4-Flash-Vision-Exp, became available on the DeepSeek API platform. The source explicitly identifies this release as an experimental model. To access the model programmatically, developers must set the model parameter to 'deepseek-v4-flash-vision-exp' within their API requests. The model is designed to accept both text and image inputs while producing text outputs. While a Hugging Face repository exists for the model, the primary documented access method in the release notes is through the DeepSeek API.
Reported Benchmark Performance
The official changelog lists specific numerical scores for several public benchmark sets. These figures represent the performance recorded during the testing phase described in the source. For Terminal Bench 2.1, the score is 83.9. NL2Repo achieved a score of 57.7, while DeepSWE recorded 59.3. DSBench-Hard shows a score of 63.6. In the AutomationBench (Public) category, the score is 25.7. ApexBench (Pass@1) reports a score of 36.5, and Agents' Last Exam records 27.3. Chartography achieved 64.3, and ZeroBench (Pass@5) reported a score of 35.0. A footnote in the source indicates that for Code Agent text tasks in these public benchmark sets, DeepSeek family models were tested using the DeepSeek Harness minimal configuration. These numbers are static data points from the release documentation and do not constitute a comparative ranking against other vendors or previous iterations without additional external data.
- Terminal Bench 2.1: 83.9
- NL2Repo: 57.7
- DeepSWE: 59.3
- DSBench-Hard: 63.6
- AutomationBench (Public): 25.7
- ApexBench (Pass@1): 36.5
- Agents' Last Exam: 27.3
- Chartography: 64.3
Evaluation Procedure and Unknowns
Engineers evaluating this experimental model should design tests based on their specific application requirements rather than relying solely on the published aggregate scores. A practical evaluation procedure involves creating a controlled test set of images and text prompts relevant to the target use case. Developers should measure the model's output accuracy against ground truth data for these specific inputs. They can compare the results against baseline models or previous internal standards to validate if the multimodal capabilities meet project needs. It is important to note that the source does not specify the context window size, maximum output tokens, or pricing structure. Additionally, while open weights are indicated as true in the metadata, the release notes focus on API availability. Users should verify local installation requirements separately if they intend to deploy outside the API environment, as the provided text does not detail deployment architecture or migration constraints.
Primary sources
Frequently asked questions
How do I access the DeepSeek-V4-Flash-Vision-Exp model?
According to the release notes dated August 21, 2026, the model is available on the DeepSeek API platform. You can access it by setting the model identifier to 'deepseek-v4-flash-vision-exp' in your API request parameters.
What are the input and output modalities for this model?
The model accepts text and image inputs. Its output modality is text. This makes it suitable for tasks requiring visual understanding combined with textual processing.